11:11

Trees are a sea in another sense.

[11:11 @T486912Los]

@Wangxi07 那什么不一般？

[11:11 @T486912Los]

昨天喝酒喝晕脱了衣服在被窝里视频。

迷迷糊糊，眼皮睁不开，当大脑一直在回答问题，黏黏糊糊。

[11:11 @T486912Los]

@liweiOrs2 也属于前排了

[11:11 @T486912Los]

@Wangxi07 🦴配酒；干杯，已经开喝

[Orange AI @oran_ge]

AI 行业发展到了新阶段
最近海内外都开始有一些反思的声音出现了：
1. model+harness才是产品（而不会吃掉一切）
2. 完全自动化就是在骗自己，珍贵的东西都无法完全自动化，离不开人的参与
3. 慢工出细活的耐心，是AI时代的新奢侈品，欲速则不达，慢就是快
4. AI 的成本支出很大，算 ROI 有时候不如人

[11:11 @T486912Los]

好消息是，有办法缓解。

专家建议把 AI 当“热情但有点天真的助手”：不要直接让它输出最终答案，而是每一步都自己先想，再用 AI 验证或补充；或者让 AI 反问你，而不是直接给答案。

你现在写文章、做方案，还敢完全不用 AI 吗？
或者你已经感受到自己变“懒”了？

https://t.co/t1W96OdfM4

[11:11 @T486912Los]

研究者把这叫“认知卸载”（cognitive offloading）。

把简单计算卸载给计算器没问题，但把写作、问题解决这种复杂思维过程全交给 AI，就可能形成恶性循环：越依赖越不会思考，越不会思考就越依赖。

有人直言：“离开 AI 我都不知道怎么解决问题了。”长期下去，企业和个人竞争力都可能受影响。

[Charly Wargnier ♨️ @DataChaz]

🚨 @AntLingAGI HAS OFFICIALLY SET A NEW STANDARD FOR OPEN-SOURCE AI

They just dropped Ring-2.6-1T, a trillion-parameter model built specifically to tackle long-horizon workflows 🔥

The benchmarks show exactly why it leads the open-source space:

→ Tau2-Bench (95.32): Navigates complex API tasks effortlessly.
→ ClawEval (63.82): Demonstrates competitive performance compared to models in the same category.
→ GPQA Diamond (88.27): Highlights strong scientific knowledge comprehension and advanced reasoning capabilities.

The takeaway for builders is simple:

We no longer need to pay premium rates for elite code execution.

In the demo below, I used Ring-2.6-1T exclusively to debug some of my 🐍 Python apps.

Debugging at scale in prod usually requires:

✦ Navigating multiple directories
✦ Evaluating shell outputs
✦ Orchestrating read/write permissions without breaking the main branch

This new OSS model handles that massive context load flawlessly.

Running it directly inside my IDE, watch the model autonomously read a broken script, locate a failing regex pattern in the server logs, and test its exact-match patch in real time 👀↓

[Siraj Raval @sirajraval]

Ant Group dropped a 100B-param model, Ling-2.6-Flash, that is near-frontier on benchmarks, fully open-weights, and 1/6 the price of GPT-5.5.

Most labs are racing on parameter count. Ling-2.6-Flash is racing on tokens-per-task. Same answer, fewer reasoning steps, lower bill. That is the actual axis customers care about once you are past the demo.

Ran it on a real workflow this week. Same coding task GPT-5.5 takes 4,200 tokens to solve, Ling-2.6-Flash hits in 1,900. At $0.15 / $0.40 per million in/out, that compounds fast on production traffic.

OpenRouter has it live: https://t.co/mdGH2sJsJo. No waitlist, no API key gymnastics. Open weights are also on HuggingFace if you want to self-host.

My full breakdown is below. The thesis: efficiency is the new IQ. Any lab that does not ship a tokens-per-task chart in their next launch is getting lapped by the ones that do.

[11:11 @T486912Los]

连库克都退休了，我还没退休

[汗青 HQ @hq4ai]

我们首次与理想汽车合作，用AI李想整了段活儿，看数据是要爆一下了。
Yuri也正式开启和理想的合作，这次先是音乐。
23年开始我们用AI尝试做“人”，庆幸一路的累积都有价值。

[11:11 @T486912Los]

@4BwNkO18LybOp8g 嗯哼

[11:11 @T486912Los]

@kkluaaaa 好想知道这件衣服的链接